You're on your third attempt. You tilt the phone at the stars, hold your breath, and wait. And then the preview resolves into something that shouldn't be possible: Milky Way, foreground trees, actual texture in the dark. You didn't touch a single setting. The camera just decided.
That's not magic. It's a specific set of tradeoffs, and once you see the machinery underneath, you'll understand why some shots stun you and others look like gray soup.
The problem with treating darkness as one thing
Your phone's sensor captures photons. In a lit room, there are plenty. Outside at midnight, there are almost none, and the ones that do arrive carry wildly different information: a faint smear of the galactic core, the warm glow of a distant town bleeding up from the horizon, the cold silhouette of a pine tree. Exposing for all of them equally would wash out the brightest and bury the dimmest.
So the phone doesn't expose for the scene. It exposes for zones within it.
Every camera has done some version of metering for decades. A phone running night mode, though, goes several steps further. It captures a burst, typically eight to thirty frames, each lasting a fraction of a second, and the image signal processor analyses what's in each one before merging them. Google's Night Sight, Apple's Night mode, and Samsung's Expert RAW all use variants of this multi-frame approach. The interesting part is what happens before the merge: the algorithm identifies which regions are information-rich and which are just noise.
Bright pinpricks against a dark field, like stars, register as high-frequency detail with low surrounding signal. The processor treats those differently from a lit window or a streetlamp, which carry high signal but little detail worth preserving. A gradient like the Milky Way core sits in a third category: medium signal, high spatial variance, which roughly means it's worth pulling more exposure toward.
The phone is essentially asking: where is there something to recover, and where is there just darkness?
A worked example that makes this concrete
Priya and Marcus both shoot the same hillside on the same night with phones from the same generation. One compositional difference separates them.
Priya puts the horizon at the lower third of the frame. The algorithm reads the upper two-thirds as star field, detects the galactic smear, and allocates longer effective exposure to that zone during frame merging. The horizon gets slightly compressed because the town glow would otherwise blow out. Her foreground is dark but textured.
Marcus puts the horizon dead center. The glow becomes the meter anchor. Stars above get comparatively less exposure time in the merge, so his sky looks flatter and the stars dimmer, though his horizon detail is cleaner.
Same hardware. Same night. The algorithm did exactly what it was told, both times. Framing is an instruction to the processor, not just a choice about aesthetics.
What people consistently misread about this
The common assumption is that night mode just brightens the image. It doesn't, not exactly. Brightening uniformly would amplify noise everywhere, and your sky would look like someone spilled salt on wet concrete.
What the processor actually does is apply differential noise reduction. Regions flagged as information-poor get heavy smoothing, which kills noise but also kills any faint detail that was there. Regions flagged as information-rich get lighter smoothing and more aggressive frame-stacking to pull signal out of the noise floor. The star field and the Milky Way core get treated like fine print worth preserving. The flat black between constellations gets smoothed nearly to nothing.
This is why a cropped phone shot of the Andromeda Galaxy can look surprisingly sharp at the core and then go oddly plastic in the surrounding sky. You're seeing the boundary between two treatment zones. It's a seam, and once you know to look for it, you'll see it everywhere.
Motion, by the way, doesn't just degrade night sky shots. It destroys them. If the phone shifts during the burst, star positions change between frames. The alignment algorithm either fails and produces trails, or, worse, it identifies the stars as inconsistent signal across frames and smooths them away entirely. The thing it was working hardest to preserve gets reclassified as a problem.
So ask yourself the obvious question: if the algorithm is this sensitive to movement, why are you still trying to handhold a six-second exposure?
A steady hand buys you almost nothing here. The physics are just against you. A wall, a fence post, a five-dollar tripod mount from a convenience store rack, any of those buys you everything, because what you're really doing is giving the algorithm the stable, consistent signal it needs to do its job properly.
The gear people obsess over matters far less than that one dull, unsexy variable. Stability isn't a tip. It's the whole game.